Anthropic Co-founder Warns AI Could Achieve Self-Improvement by 2028

According to Jack Clark's speech at Oxford University, Anthropic's co-founder warned that artificial intelligence could achieve recursive self-improvement—the ability to autonomously upgrade itself and create more advanced versions—potentially by 2028 or earlier. Clark acknowledged that Anthropic itself has significantly underestimated the pace of AI advancement and is underprepared for risks posed by cutting-edge models.

Clark cited Mythos, an internal model completed in April 2026, as evidence of these risks. The model possesses capabilities comparable to nation-state-level cyberattack weapons. Due to its destructive potential, Anthropic has indefinitely banned public release of Mythos, restricting access to a limited number of institutions for software vulnerability research only.